I noted this in a tram-replacement experiment, with histogram code from bale suite. Nodegroup messages carrying boatload of data were arriving on the node but not being processed as all PEs were busy cycling through there own locally generated work. (Although the work generation method was broken down using CThYield() every x iterations). @nilsdeppe also seconded this request on slack.
What is happening: scheduler's function for finding next message always processes non-local (i.e. messages from others but meant for my PE) and local (from my pe) messages, so nodeGroup messages, although they can be processed by any PE, get delayed.
The request is to allow nodegroup messages to execute earlier.. somehow.